X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;ds=inline;f=qscale.c;h=8dafced51c5f49ef7802d4748cbd0ab615721c4b;hb=eaf1b308bc94f0e4b6a8d2366ca02499b5255221;hp=1b3ab57c0b9bd20a76785d66caac745b445a774c;hpb=0f5c62013c0210e1f9806342eb1271674f735d7d;p=qscale diff --git a/qscale.c b/qscale.c index 1b3ab57..8dafced 100644 --- a/qscale.c +++ b/qscale.c @@ -9,8 +9,7 @@ double sinc(double x) { - static const double eps = 2.22045e-016; - static const double cutoff = sqrt(sqrt(eps)); + static const double cutoff = 1.220703668e-4; // sqrt(sqrt(eps)) if (abs(x) < cutoff) { // For small |x|, use Taylor series instead @@ -39,7 +38,7 @@ struct pix_desc { unsigned startcoeff; }; -void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw, unsigned dstride) +void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw, unsigned sstride, unsigned dstride) { struct pix_desc *pd = (struct pix_desc *)malloc(nw * sizeof(struct pix_desc)); int size_coeffs = 8; @@ -87,7 +86,7 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw size_coeffs <<= 1; coeffs = (float *)realloc(coeffs, size_coeffs * sizeof(float)); } - + coeffs[num_coeffs++] = f; sum += f; } @@ -96,36 +95,41 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw coeffs[pd[x].startcoeff + sx - start] /= sum; } } - + for (y = 0; y < h; ++y) { - float *sptr = pix + y*w; + float *sptr = pix + y*sstride; unsigned char *dptr = npix + y*dstride; unsigned char ch; for (x = 0; x < nw; ++x) { + int result; float acc; + long tmp; static const float low = 0.0, high = 255.0; asm ( - "pxor %0, %0 \n" - "xor %%eax, %%eax \n" + "pxor %1, %1 \n" + "xor %2, %2 \n" ".lbl2: \n" - "movups (%2,%%eax),%%xmm1 \n" - "movups (%1,%%eax),%%xmm2 \n" + "movups (%4,%2),%%xmm1 \n" + "movups (%3,%2),%%xmm2 \n" "mulps %%xmm2,%%xmm1 \n" - "addps %%xmm1,%0 \n" - "addl $16,%%eax \n" - "dec %3 \n" + "addps %%xmm1,%1 \n" + "add $16,%2 \n" + "dec %5 \n" "jnz .lbl2 \n" - "haddps %0,%0 \n" - "haddps %0,%0 \n" - "maxss %4,%0 \n" - "minss %5,%0 \n" - : "=x" (acc) + "haddps %1,%1 \n" + "haddps %1,%1 \n" + "maxss %6,%1 \n" + "minss %7,%1 \n" + "cvtss2si %1,%0 \n" + : "=r" (result), + "=&x" (acc), + "=&r" (tmp) : "r" (&coeffs[pd[x].startcoeff]), "r" (&sptr[pd[x].start]), "r" ((pd[x].end - pd[x].start + 1)/4), "m" (low), "m" (high) - : "eax", "xmm1", "xmm2" + : "xmm1", "xmm2" ); #if 0 @@ -137,7 +141,7 @@ void hscale(float *pix, unsigned char *npix, unsigned w, unsigned h, unsigned nw ch = (unsigned char)acc; *dptr++ = ch; #endif - *dptr++ = (unsigned char)acc; + *dptr++ = (unsigned char)result; } ch = dptr[-1]; for ( ; x < dstride; ++x) { @@ -221,83 +225,83 @@ void vscale(unsigned char *pix, float *npix, unsigned w, unsigned h, unsigned nh */ asm ( /* clear */ - "pxor %%xmm0, %%xmm0 \n" - "pxor %%xmm1, %%xmm1 \n" - "pxor %%xmm2, %%xmm2 \n" - "pxor %%xmm3, %%xmm3 \n" + "pxor %%xmm0, %%xmm0 \n" + "pxor %%xmm1, %%xmm1 \n" + "pxor %%xmm2, %%xmm2 \n" + "pxor %%xmm3, %%xmm3 \n" /* main loop */ - ".lbl: \n" + ".lbl: \n" /* a zero is useful during unpacking */ - "pxor %%xmm4, %%xmm4 \n" + "pxor %%xmm4, %%xmm4 \n" /* fetch all 16 source bytes */ - "movups (%0), %%xmm5 \n" - "prefetcht0 (%0,%3,4) \n" + "movups (%0), %%xmm5 \n" + "prefetcht0 (%0,%3,4) \n" /* unpack into words (xmm5, xmm7) */ - "movaps %%xmm5, %%xmm7 \n" - "punpcklbw %%xmm4, %%xmm5 \n" - "punpckhbw %%xmm4, %%xmm7 \n" + "movaps %%xmm5, %%xmm7 \n" + "punpcklbw %%xmm4, %%xmm5 \n" + "punpckhbw %%xmm4, %%xmm7 \n" /* unpack xmm5 into dwords (xmm5, xmm6) */ - "movaps %%xmm5, %%xmm6 \n" - "punpcklwd %%xmm4, %%xmm5 \n" - "punpckhwd %%xmm4, %%xmm6 \n" + "movaps %%xmm5, %%xmm6 \n" + "punpcklwd %%xmm4, %%xmm5 \n" + "punpckhwd %%xmm4, %%xmm6 \n" /* convert xmm5, xmm6 to floats */ - "cvtdq2ps %%xmm5, %%xmm5 \n" - "cvtdq2ps %%xmm6, %%xmm6 \n" + "cvtdq2ps %%xmm5, %%xmm5 \n" + "cvtdq2ps %%xmm6, %%xmm6 \n" /* fetch the coefficient */ - "movss (%2), %%xmm4 \n" - "shufps $0x0, %%xmm4, %%xmm4 \n" + "movss (%2), %%xmm4 \n" + "shufps $0x0, %%xmm4, %%xmm4 \n" /* do the muls for xmm5 and xmm6 */ - "mulps %%xmm4, %%xmm5 \n" - "mulps %%xmm4, %%xmm6 \n" - "addps %%xmm5, %%xmm0 \n" - "addps %%xmm6, %%xmm1 \n" + "mulps %%xmm4, %%xmm5 \n" + "mulps %%xmm4, %%xmm6 \n" + "addps %%xmm5, %%xmm0 \n" + "addps %%xmm6, %%xmm1 \n" /* get the zero back again */ - "pxor %%xmm4, %%xmm4 \n" + "pxor %%xmm4, %%xmm4 \n" /* unpack xmm7 into dwords (xmm7, xmm6) */ - "movaps %%xmm7, %%xmm6 \n" - "punpcklwd %%xmm4, %%xmm7 \n" - "punpckhwd %%xmm4, %%xmm6 \n" + "movaps %%xmm7, %%xmm6 \n" + "punpcklwd %%xmm4, %%xmm7 \n" + "punpckhwd %%xmm4, %%xmm6 \n" /* convert xmm7, xmm6 to floats */ - "cvtdq2ps %%xmm7, %%xmm7 \n" - "cvtdq2ps %%xmm6, %%xmm6 \n" + "cvtdq2ps %%xmm7, %%xmm7 \n" + "cvtdq2ps %%xmm6, %%xmm6 \n" /* fetch the coefficient */ - "movss (%2), %%xmm4 \n" - "shufps $0x0, %%xmm4, %%xmm4 \n" + "movss (%2), %%xmm4 \n" + "shufps $0x0, %%xmm4, %%xmm4 \n" /* do the second set of muls */ - "mulps %%xmm4, %%xmm7 \n" - "mulps %%xmm4, %%xmm6 \n" - "addps %%xmm7, %%xmm2 \n" - "addps %%xmm6, %%xmm3 \n" + "mulps %%xmm4, %%xmm7 \n" + "mulps %%xmm4, %%xmm6 \n" + "addps %%xmm7, %%xmm2 \n" + "addps %%xmm6, %%xmm3 \n" /* move along, and loop */ - "add $4, %2 \n" - "add %3, %0 \n" - "dec %1 \n" - "jnz .lbl \n" + "add $4, %2 \n" + "add %3, %0 \n" + "dec %1 \n" + "jnz .lbl \n" /* store the values */ - "movaps %%xmm0, (%4) \n" - "movaps %%xmm1, 16(%4) \n" - "movaps %%xmm2, 32(%4) \n" - "movaps %%xmm3, 48(%4) \n" + "movaps %%xmm0, (%4) \n" + "movaps %%xmm1, 16(%4) \n" + "movaps %%xmm2, 32(%4) \n" + "movaps %%xmm3, 48(%4) \n" : : "r" (&sptr[pd[y].start * w]), /* 0: srcptr base */ "r" (pd[y].end - pd[y].start + 1), /* 1: filter len */ "r" (&coeffs[pd[y].startcoeff]), /* 2: coeffs base */ - "r" (w), /* 3: stride */ + "r" ((long)w), /* 3: stride */ "r" (dptr) /* 4: dstptr base */ : "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" ); @@ -353,6 +357,15 @@ int main(int argc, char **argv) dinfo.raw_data_out = TRUE; jpeg_start_decompress(&dinfo); + unsigned w0 = dinfo.image_width * dinfo.comp_info[0].h_samp_factor / dinfo.max_h_samp_factor; + unsigned h0 = dinfo.image_height * dinfo.comp_info[0].v_samp_factor / dinfo.max_v_samp_factor; + + unsigned w1 = dinfo.image_width * dinfo.comp_info[1].h_samp_factor / dinfo.max_h_samp_factor; + unsigned h1 = dinfo.image_height * dinfo.comp_info[1].v_samp_factor / dinfo.max_v_samp_factor; + + unsigned w2 = dinfo.image_width * dinfo.comp_info[2].h_samp_factor / dinfo.max_h_samp_factor; + unsigned h2 = dinfo.image_height * dinfo.comp_info[2].v_samp_factor / dinfo.max_v_samp_factor; + fprintf(stderr, "Scaling using Lanczos filter:\n"); fprintf(stderr, " Y component: %ux%u -> %ux%u\n", dinfo.comp_info[0].width_in_blocks * DCTSIZE, dinfo.comp_info[0].height_in_blocks * DCTSIZE, nw0, nh0); fprintf(stderr, " Cb component: %ux%u -> %ux%u\n", dinfo.comp_info[1].width_in_blocks * DCTSIZE, dinfo.comp_info[1].height_in_blocks * DCTSIZE, nw1, nh1); @@ -388,23 +401,23 @@ int main(int argc, char **argv) { float *npix = (float*)memalign(16, dinfo.comp_info[0].width_in_blocks * DCTSIZE * nh0 * sizeof(float)); - vscale(data_y, npix, dinfo.comp_info[0].width_in_blocks * DCTSIZE, dinfo.comp_info[0].height_in_blocks * DCTSIZE, nh0, dinfo.comp_info[0].width_in_blocks * DCTSIZE); + vscale(data_y, npix, dinfo.comp_info[0].width_in_blocks * DCTSIZE, h0, nh0, dinfo.comp_info[0].width_in_blocks * DCTSIZE); data_ny = (unsigned char *)malloc(nw0 * stride0); - hscale(npix, data_ny, dinfo.comp_info[0].width_in_blocks * DCTSIZE, nh0, nw0, stride0); + hscale(npix, data_ny, w0, nh0, nw0, dinfo.comp_info[0].width_in_blocks * DCTSIZE, stride0); free(npix); } { float *npix = (float*)memalign(16, dinfo.comp_info[1].width_in_blocks * DCTSIZE * nh1 * sizeof(float)); - vscale(data_cr, npix, dinfo.comp_info[1].width_in_blocks * DCTSIZE, dinfo.comp_info[1].height_in_blocks * DCTSIZE, nh1, dinfo.comp_info[1].width_in_blocks * DCTSIZE); + vscale(data_cr, npix, dinfo.comp_info[1].width_in_blocks * DCTSIZE, h1, nh1, dinfo.comp_info[1].width_in_blocks * DCTSIZE); data_ncr = (unsigned char *)malloc(nw1 * stride1); - hscale(npix, data_ncr, dinfo.comp_info[1].width_in_blocks * DCTSIZE, nh1, nw1, stride1); + hscale(npix, data_ncr, w1, nh1, nw1, dinfo.comp_info[1].width_in_blocks * DCTSIZE, stride1); free(npix); } { float *npix = (float*)memalign(16, dinfo.comp_info[2].width_in_blocks * DCTSIZE * nh2 * sizeof(float)); - vscale(data_cb, npix, dinfo.comp_info[2].width_in_blocks * DCTSIZE, dinfo.comp_info[2].height_in_blocks * DCTSIZE, nh2, dinfo.comp_info[2].width_in_blocks * DCTSIZE); + vscale(data_cb, npix, dinfo.comp_info[2].width_in_blocks * DCTSIZE, h2, nh2, dinfo.comp_info[2].width_in_blocks * DCTSIZE); data_ncb = (unsigned char *)malloc(nw2 * stride2); - hscale(npix, data_ncb, dinfo.comp_info[2].width_in_blocks * DCTSIZE, nh2, nw2, stride2); + hscale(npix, data_ncb, w2, nh2, nw2, dinfo.comp_info[2].width_in_blocks * DCTSIZE, stride2); free(npix); } jpeg_destroy_decompress(&dinfo);